[R-sig-ME] poisson GLMER with identity link

[Ben Bolker]

  I don't think there is a way to do this without hacking.
  I don't know what glm.fit does, internally, to address this situation
-- certainly one gets a lot of warnings, but it does come up with a
reasonably sensible answer (once reasonable starting values have been
specified).
  If I were going to hack this I would go into lmer.c , find the
location where the warning occurred, and change the code to set the
offending value to a feasible (non-negative) value instead (giving a
warning).  Slightly more conservatively you could set a threshold below
which the value was adjusted with a warning and above which an error
occurred.

> x = runif(100)
> y = rpois(100,4*x)
> glm(y~x,family=poisson(link="identity"))
Error: no valid set of coefficients has been found: please supply
starting values
In addition: Warning message:
In log(ifelse(y == 0, 1, y/mu)) : NaNs produced
> glm(y~x,family=poisson(link="identity"),start=c(0.1,4))

Call:  glm(formula = y ~ x, family = poisson(link = "identity"), start =
c(0.1,      4))

Coefficients:
(Intercept)            x
   -0.01669      4.28255

Degrees of Freedom: 99 Total (i.e. Null);  98 Residual
Null Deviance:	    176.8
Residual Deviance: 87.02 	AIC: 295
There were 15 warnings (use warnings() to see them)
> warnings()
Warning messages:
1: step size truncated: out of bounds
2: step size truncated: out of bounds
3: step size truncated: out of bounds
4: step size truncated: out of bounds
5: step size truncated: out of bounds
6: step size truncated: out of bounds
7: step size truncated: out of bounds
8: step size truncated: out of bounds
9: step size truncated: out of bounds
10: step size truncated: out of bounds
11: step size truncated: out of bounds
12: step size truncated: out of bounds
13: step size truncated: out of bounds
14: In glm.fit(x = X, y = Y, weights = weights, start = start,  ... :
  algorithm stopped at boundary value
15: In glm.fit(x = X, y = Y, weights = weights, start = start,  ... :
  fitted rates numerically 0 occurred


Tim Carnus wrote:
> Dear list,
> 
> Just adding an example of the model I am fitting, and session info.
> 
> model<-glmer(count~-1+X1+X2+X3+X4+(1|plot),poisson(link='identity'),
> REML=TRUE)
> 
> where I am interested in the additive effects of X1-X4 on the count per
> plot response. The random effect is simply there to account for the
> repeated measures taken over time on each plot.
> 
> R version 2.10.1 (2009-12-14) 
> i486-pc-linux-gnu 
> 
> 
> 
> On Mon, 2010-03-29 at 17:45 +0100, Tim Carnus wrote:
>> Dear list,
>>
>> I am trying to fit a number of GLMERs to count data with an additive
>> model (in the predictors) that requires the use of the identity link
>> function. For about half of my response variables this causes no
>> problems. However in a number of cases the model fitting runs into
>> problems with regards estimation of negative mean (for e.g. the error
>> message in mer_finalize: mu[i] must be positive: mu = 1.76267e-312, i =
>> 13075456). As far as I understand this is well known and documented, and
>> guarding against that possibility is necessary, and built in to say the
>> glm() function.
>>
>> My question then is, how can I do this with lmer? (ie how can I specify
>> the constraints necessary to fit these types of models, if at all
>> possible)
>>
>> Best regards,
>>
>> Tim Carnus
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models


-- 
Ben Bolker
Associate professor, Biology Dep't, Univ. of Florida
bolker at ufl.edu / people.biology.ufl.edu/bolker
GPG key: people.biology.ufl.edu/bolker/benbolker-publickey.asc